Improving Reduced Reference Image Quality Assessment Methods By Using Color Information

Abstract

In real-world applications, images and videos are often acquired and displayed in color. To assess the quality of these images, most of the methods have been developed in a grayscale level without considering the color information. Among the methods that remain less explored in color, one can find reduced reference image quality assessment (RRIQA) methods and especially those based on natural scenes statistics (NSS). Motivated by the close relationship that lies between inter/intra-color components perception and statistics, this paper proposes a new framework to study the impact of color information on RRIQA methods and more specifically the NSS based ones. For this purpose, the inquiry investigates how each information (luminance, chrominance) influences the quality assessment process. Then, it also considers whether the combination of these components can improve the quality prediction scores. The deployment of this framework is closely related to the choice of quality methods and perceptual color spaces. Thus, four of the most influential RRIQA based NSS methods have been intuitively extended to color. Furthermore, YCbCr and CIELAB color spaces are selected thanks to their usefulness to separate chrominance and luminance information. On an experimental level, these methods are implemented on TID2013 benchmark, which offers a wide range of color specified distortion types. The obtained results showcase how color information can improve quality scores.